SomewhatLikely

Cool idea. This might be a time where an option to have an alphabetically ordered keyboard would be useful.

By the way, I like testing AIs on new games as they aren't in the training data and gpt5.6 sol high thinking was able to get from my second guess result to the correct answer. One letter was already locked and four more were down to a five letterish range. Still impressive as I remember models up until about the second generation of reasoning models really struggling with games like wordle.

Checking my history I searched ["chaos goblin" chatgpt] on March 6th after seeing too many goblins and gremlins and didn't find anyone talking about it then. I did have the nerdy personality turned on and in my testing of Chatgpt 5.5 I did notice the nerdy personality was gone because some responses were not considering as many plausible interpretations or covering as many useful answers as the response recorded for 5.4. Rather than having the LLM guess the most plausible interpretation and focus on the most likely answer I prefer a more well-rounded response and if I want less I'll scan. Anyway, after seeing the personality was gone I just added a custom instruction to take on a nerdy persona and got back my desired behavior. But also the gremlins and goblins are back so I don't think their mitigation is strong enough to overcome the personality tuning.

There would still be a point to encourage better predictions from the public information through better modeling. We aren't always using the optimal models to predict. One example: LLMs are "just" predicting the next token given the public information of the tokens that came before, but they work considerably better at making that prediction than the models that came before them.

If I'm understanding correctly, this might be the first example of just-in-time agentic coding that I've come across.

I've seen it so this too. I had it keeping a running tally over many turns and occasionally it would say something like: "... bringing the total to 304.. 306, no 303. Haha, just kidding I know it's really 310." With the last number being the right one. I'm curious if it's an organic behavior or a taught one. It could be self learned through reinforcement learning, a way to correct itself since it doesn't have access to a backspace key.

The default outputs are considerably shorter even in thinking mode. Something that helped me get the thinking mode back to an acceptable state was to switch to the Nerd personality and in the traits customization setting tell it to be complete and add extra relevant details. With those additions it compares favorably to o3 on my recent chat history and even improved some cases. I prefer to scan a longer output than have the LLM guess what to omit. But I know many people have complained about verbosity so I can understand why they may have moved to less verbiage.

Nobody's preventing them from rendering it and refining. That's certainly what we'd expect an AGI to do.

It's likely that the commenter has read less than 5 million posts worth of text though. So perhaps this still points to a lack of diversity in content.

What was the alternative you went with?

I saw a very similar timely appeal here on Hacker News a few years ago and taught my son with this book at the age of 4. It has become my go-to comparison when prompting chat bots on what I want in a teaching material for other subjects. I listened to the entire article posted here and it makes me wonder if schools are getting something as foundational as reading wrong how can we trust the attention to research on anything else they're teaching? Don't get me wrong, I'm not going to pull my kid out of school but I'll dig a little deeper into how well he's learning. For math, we've been doing the Beast Academy books. It has gone... Okay. I like that they approach problems from many different ways which simulate the many different ways math is hidden in our interactions with the world. For my younger son I've recently started Teaching Your Child... because of how well it went for his brother but for math I may try something else to have a new data point. Something that occurred to me listening to the article is I wonder if certain skills are learned much faster with one on one instruction like the book has you do. Our schools pretty much never teach that way out of efficiency, though home schools often do. It may not be true for most subjects though or home school students would be so far ahead by college and that's not the impression I have.

It's pretty damn capital intensive to be a productive farmer today. That said, AI will likely, hopefully, get cheaper over time.

You could probably modify the metric to addressed comments per 1000 lines of code.

My first thought upon seeing the prompt:

    If you would build an in-memory cache, how would you do it?

    It should have good performance and be able to hold many entries. 
    Reads are more common than writes. I know how I would do it already, 
    but I’m curious about your approach.

Was to add this requirement since it comes up so often:

    Let's assume that keys accessed follow a power law, so some keys get 
    accessed very frequently and we would like them to have the fastest 
    retrieval of all.

I'm not sure if there are any efficient tweaks to hash tables or b-trees that might help with this additional requirement. Obviously we could make a hash table take way more space than needed to reduce collisions, but with a decent load factor is the answer to just swap frequently accessed keys to the beginning of their probe chain? How do we know it's frequently accessed? Count-Min sketch?

Even with that tweak, the hottest keys will still be scattered around memory. Wouldn't it be best if their entries could fit into fewer pages? So, maybe a much smaller "hot" table containing say the 1,000 most accessed keys. We still want a high load factor to maximize the use of cache pages so perhaps perfect hashing?

Where I thought this might be going from the first paragraph:

Negative numbers are sometimes represented with parentheses: (234.58)

Tables sometimes tell you in the description that all numbers in are in 1000's or millions.

The dollar sign is used by many currencies, including in Australia and Canada.

I'd probably look around for some other gotchas. Here's one page on prices in general: https://gist.github.com/rgs/6509585 but interestingly doesn't quite cover the OP's problem or the ones I brought up, though the use cases are slightly different.

Something worth mentioning is that if your vectors all have the same length then cosine similarity and Euclidean distance will order most (all?) neighbors in the same order. Think of your query vector as a point on a unit sphere. The Euclidean distance to a neighbor will be a chord from the query point to the neighbor. Just as with the angle between the query-to-origin and the neighbor-to-origin vectors, the farther you move the neighbor from the query point on the surface of the sphere, the longer the chord between those points gets too.

EDIT: Here's a better treatment, and it is the case that they give the exact same orderings: https://ajayp.app/posts/2020/05/relationship-between-cosine-...

This feels similar to those adversarial examples that first came out that were very tuned for a specific image recognizer. I haven't followed the research but I know they had some very limited success to getting it to work in the real world. I'm not sure if they ever worked across different models though.

The paper claims there is literature with more success for LLMs:

   Large language models have been shown to be vulnerable to adversarial
   attacks, in which attackers introduce maliciously crafted token sequences
   into the input prompt to circumvent the model’s safety mechanisms and 
   generate a harmful response [1, 14].

Can I recommend you provide some cost estimates next to the examples for using your own key? I tried a few custom extractions and then checked my usage dashboard and it was already over $2.

https://arxiv.org/abs/2006.16236 Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention

Video games are gonna be wild in the near future. You could have one person talking to a model producing something that's on par with a AAA title from today. Imagine the 2d sidescroller boom on Steam but with immersive photorealistic 3d games with hyper-realistic physics (water flow, fire that spreads, tornados) and full deformability and buildability because the model is pretrained with real world videos. Your game is just a "style" that tweaks some priors on look, settings, and story.

I've switched to LLM's for explaining commands. Here's ChatGPT4o:

The command tar xvzf 1.zip appears to be trying to extract a compressed archive file, but there's a mismatch between the file extension (1.zip) and the command options used for tar. Let me break it down:

tar: This is the command-line utility for working with tar archives (tape archives). It can be used to create, extract, and manipulate tar files.

x: This option tells tar to extract files from an archive.

v: This stands for "verbose", which means that tar will list the files being extracted to the terminal.

z: This option tells tar to handle a compressed archive using gzip. It is typically used with .tar.gz or .tgz files.

f: This tells tar that the next argument is the name of the archive file.

However, the file 1.zip is a ZIP file, not a tar.gz file. ZIP files are handled by the unzip command, not tar. Using tar xvzf on a .zip file would result in an error, as tar expects a .tar.gz file when using the z option.

To properly extract a .zip file, you should use:

unzip 1.zip

This command will extract the contents of the ZIP file.

I suspect there may be a flaw in the visualization where smaller dots are drawn with disproportionately more pixels than bigger dots. Just looking at how much green is on the map is way more than numbers I can find support for. For example, spain visually looks like more than 90% solar/wind but in reality it appears to only be about 60% [ https://www.reuters.com/business/energy/renewables-produce-a... ]

Both of your references use RRF=1/(60+Rank)

So I'm not sure why the article uses 1/Rank alone. Did you test both and find that the smoothing didn't help? My understanding is that it has been pretty important for the best results.

I wonder if this would work in a rice cooker or if there's something equivalent. If I don't wash the rice beforehand it will tend to bubble starchy bubbles through the top that makes a mess

What would this do that a minimum wage wouldn't do?

A chart showing BMI vs altitude could help cut through individual bases differences. If there is a trend downward it would show a dose-response effect. I didn't find such a chart in the paper. Instead they have a chart showing how many data points were available by altitude.

Thanks, this gave some good insight to GPT4. If I provide the entire Wikipedia page contents but blank out the movie name and director name it can't recall it. https://chat.openai.com/share/c499e163-3745-48c3-b00e-11ea42...

However, if I add the director it gets it right: https://chat.openai.com/share/a602b3b0-5c17-4b4d-bed8-124197...

If I only give it 1980s film and the director's name it can still get it. https://chat.openai.com/share/d6cf396b-3199-4c80-84b9-d41d23...

So it's clearly not able to look this movie up semantically and needs a strong key like the director's name.

EDIT: Digging deeper it's clear the model only has a very foggy idea of what the movie is about: https://chat.openai.com/share/d0701f53-1250-421e-aa4b-dc8156... People have described these types of outputs as the text equivalent of a highly compressed JPEG, which seems to fit well with what's going on here. It gets some of the top level elements right and kind of remembers there's some kind of vehicle that's important but it has forgotten all the details, even the date the movie was released. But unlike a human who might indicate their fuzziness (was it a helicopter or a submarine?), GPT4 gladly pretends like it knows what it's talking about. I think it's likely a solvable problem, the model probably has the information to know when it's confident and when it's in a fuzzy JPEG region but the current alignment isn't doing a great job of surfacing that.

That input might be too far outside of the training conditions for it to believe the sensors. It would probably assume those sensors you perturbed are faulty and ignore them.